Abstract Background The recurrence of colorectal cancer (CRC) after surgery poses a significant threat to patient recovery, but current prediction methods lack sensitivity. In this study, we aimed to develop a radiomics-based model for recurrence prediction and to investigate its association with the tumor immune microenvironment. Methods A retrospectively cohort study was conducted with 197 patients from the Second Affiliated Hospital of Harbin Medical University (Harbin, China; training set) and 130 patients from Shandong First Medical University Affiliated Central Hospital (Jinan, China; validation set). Clinical characteristics and blood biomarkers (alpha-fetoprotein, carcinoembryonic antigen, and cancer antigen 19-9) were compared between the recurrence and nonrecurrence groups, and magnetic resonance imaging (MRI)-derived radiomics features were extracted by using 3D Slicer and PyRadiomics. After this, LASSO regression was used to identify key features, and univariate analysis was used to retain statistically significant predictors. A multivariate logistic regression model was then developed to predict recurrence, and the relationship between radiomics features and immune cells was subsequently explored by using a CRC mouse model. Results There were no significant differences in the clinical characteristics between recurrence and nonrecurrence groups. The initial model, comprising 10 features, achieved an area under the curve (AUC) of 0.99 in the training set but only 0.6 in the validation set. After excluding six features, the refined model, which included original gray level co-occurrence matrix sum entropy (OGSE), log-sigma-3-0-mm-3D neighborhood gray tone difference matrix Busyness (LNB), original first order total energy (OFTE), wavelet-LLH gray level size zone matrix zone variance (WGZV), achieved an AUC of 0.98 in the validation set, outperforming traditional blood biomarkers. In a murine model, regulatory T cells exhibited a strong positive correlation with OGSE (r = 0.84, P 0.001) and moderate correlations with OFTE (r = 0.41, P = 0.035) and WGZV (r = 0.52, P = 0.007), indicating a link between radiomics features and immune cell infiltration. Conclusions MRI-derived radiomics features, particularly OGSE, LNB, OFTE, and WGZV, can effectively predict postoperative CRC recurrence. These features are correlated with the tumor immune microenvironment, supporting the use of radiomics as a noninvasive tool for tumor assessment in CRC.
Zhao et al. (Tue,) studied this question.